Clément Dumas

New paper w/@jkminder & @NeelNanda5! What do chat LLMs learn in finetuning? Anthropic introduced a tool for this: crosscoders, an SAE variant. We find key limitations of crosscoders & fix them with BatchTopK crosscoders This finds interpretable and causal chat-only features! 🧵

@snigus If you replace the half layers of Gemma 2b chat with the base model one you also get openAI #A__Other_techniques_in_the_model_diffing_toolkit" target="_blank" rel="nofollow noopener">lesswrong.com/posts/xmpauEXE…

I have a small language model and it’s been pre trained. Now I post train it to say “I’m a language model”. With no mention of openAI The trained model still ends up saying it’s a LLM made by openAI. Even tho OpenAI is never mentioned in the instruction tuning dataset, and there in fact is one sample that says “I am being developed by Anthropic” (not true)! Makes me think models saying they’re made by so and so is pretty weak evidence of copying /stealing / distillation.

@louisvarge You can also hack the built-in team feature to make this work. Was about to write a post about this but feels like channels might be cleaner?

i made a thing where now any Claude Code can send messages to any other Claude Code on my machine they can ask clarifying questions about work, or become friends

@maxsloef Welcome to the club #issuecomment-3513763128" target="_blank" rel="nofollow noopener">github.com/huggingface/pe…

the HF Llama 3 tokenizer was silently stripping spaces before punctuation on decode. every hf llama 3 model, every fine-tune, every descendant. trillions of tokens. i’m so grateful this stuff is open source — but man, the fact that only a handful of people have ever noticed this really makes you wonder how many other subtle but important bugs are just silently lurking, across all these trillions of tokens

credit to my friend nataliia for the highlights!

Read the full piece: butanium.github.io/boom-incident/

In the first thread I focused on the funny stuff but there is so much more, like some very touching moment in this piece 🥺

@thkostolansky @liminal_bardo @ESYudkowsky Me at my next RM meeting: so I did a small side project...

@Butanium_ @liminal_bardo @ESYudkowsky “i do ai research”

I asked Claude Code to "nuke" a task on my cluster, then sent "boom" >100 times in the chat. Opus 4.6 built an very flore including a NeurIPS best paper award, a @ESYudkowsky tweet thread, an EU Boom Act, half of Anthropic reacting, and much more. 🧵 (1/10)

@AdriGarriga I was literally laughing at my computer screen in the MATS office lol

@Butanium_ I'm glad you actually read Claude's replies after the booms. Seems it judged you correctly.

I should have ended with the link to the actual post: butanium.github.io/boom-incident/ or butanium.github.io/boom-incident/… for a selection of cool artifact from the conversation

One last thing in case it's not obvious, the original transcript didn't include renders of the tweets in html, just md render, e.g. "*@sama:*\n\n"we're excited [...]"\n\n*community note: \"it cannot\"*" See the full transcript here: github.com/Butanium/boom-…